Climate mode deviation correction method based on ensemble Kalman filtering and deep learning

Abstract

The invention discloses a climate mode deviation correction method based on ensemble Kalman filtering and deep learning, and the method comprises the steps: obtaining first climate historical simulation data and future climate estimation data in a research region, and obtaining historical observation data corresponding to climate variables; performing deviation correction on the first climate history simulation data point by point by adopting an ensemble filtering method; carrying out lag processing on the first climate history simulation data; constructing a long short-term memory network model; performing feature engineering processing on future climate estimation data; and performing deviation correction on future climate estimation data by using the average deviation of the Kalman gain sequence and the historical sequence of each time period in the future. According to the method, the dynamic correction capability of the ensemble Kalman filtering and the nonlinear prediction advantage of the deep learning model are combined, the correction precision of climate mode simulation data is improved, and effective deviation correction is carried out on future climate estimation data.

Description

Technical Field

[0001] The present invention relates to the technical field of environmental change impact assessment, and in particular to a climate model bias correction method based on ensemble Kalman filtering and deep learning. Background Art

[0002] The Global Climate Model (GCM) is a key tool in the field of climate and hydrological sciences for simulating and predicting changes in the Earth's climate system. It provides long-term projections from historical climate to future climate change by numerically simulating the complex interactions of the atmosphere, ocean, ice sheet, and land. However, although GCM can capture large-scale climate change trends, its output climate data usually has deviations at the regional scale. These deviations mainly come from the resolution of the model, the uncertainty of the parameterization scheme, and the imprecise input data and boundary conditions. Therefore, bias correction of GCM output data has become a key step in improving the accuracy of regional climate projections.

[0003] Existing bias correction methods include statistical-based bias correction methods and physical-based bias correction methods. Although these methods can reduce simulation biases to a certain extent, they still have limitations when dealing with complex climate data. As an advanced data assimilation method, Ensemble Kalman Filter (EnKF) can dynamically correct the model state, but it depends on observational data. The correction accuracy in the period without observational data depends on the estimation quality of the Kalman gain. Deep learning methods perform well in time series prediction and pattern recognition, and can capture complex nonlinear relationships in sequences. However, they also rely on observational data, lack physical explanations, and the generalization ability of the model needs to be studied. There is still a research gap in how to combine deep learning methods with Ensemble Kalman Filter methods with physical mechanisms to improve the bias correction accuracy of climate model simulation results. Summary of the invention

[0004] Purpose of the invention: In view of the above problems, the purpose of the present invention is to provide a climate model bias correction method based on ensemble Kalman filtering and deep learning.

[0005] Technical solution: A climate model bias correction method based on ensemble Kalman filtering and deep learning of the present invention comprises the following steps:

[0006] Step 1: Use the global climate model to obtain the first climate historical simulation data and future climate projection data in the study area, and obtain the historical observation data of the corresponding climate variables;

[0007] Step 2, based on the first climate historical simulation data and the historical observation data, the first climate historical simulation data is bias-corrected point by point using the ensemble Mann filtering method to obtain a corrected climate data sequence and Kalman gain sequence for each spatial grid point;

[0008] Step 3, performing hysteresis processing on the first climate historical simulation data to obtain second climate historical simulation data, performing feature engineering processing on the second climate historical simulation data and the Kalman gain sequence to obtain first feature engineering and target Kalman gain sequence feature engineering respectively;

[0009] Step 4, constructing a long short-term memory network model, and training the long short-term memory network model by taking the first feature engineering as an input item and the target Kalman gain sequence feature engineering as an output item;

[0010] Step 5, feature engineering is performed on the future climate prediction data to obtain a second feature engineering, and the second feature engineering is input into the trained long short-term memory network model to obtain the Kalman gain sequence of each future period;

[0011] Step 6: Based on the ensemble Kalman filter method, the future climate prediction data is bias-corrected using the average deviation of the Kalman gain sequence of each future period and the historical sequence to generate corrected future climate prediction data.

[0012] Furthermore, in step 2, the first climate historical simulation data is bias-corrected point by point using the ensemble Mann filtering method, and the corrected meteorological data sequence and Kalman gain sequence for each spatial grid point are obtained, including:

[0013] For each spatial grid point (i, j), the simulated state of the global climate model is represented by the state vector S i,j (t) indicates that the vector S i,j (t) represents the simulated value of the climate variable at time t, that is, the first climate historical simulation data, and the historical observation data of the corresponding climate variable is recorded as vector O i,j (t), calculate the deviation D between the historical observation data and the first climate historical simulation data at time t i,j (t) is:

[0014] D i,j (t) = O i,j (t)-S i,j (t)

[0015] In the formula, i and j represent the longitude and latitude of the spatial grid point respectively;

[0016] The ensemble Kalman filter algorithm advances the simulation state in time based on the physical equations of the climate model, constructs the state transfer equation and observation set with error disturbance, and obtains the predicted background field as follows:

[0017]

[0018] In the formula, ε i,j represents the simulation error, which follows a normal distribution with a mean of 0 and a variance of G, i.e., ε~N(0,G);

[0019] Calculate the mean of the background field, the calculation expression is:

[0020]

[0021] In the formula, E n represents the total number of sets m in the background field;

[0022] According to the background field and its mean, the prediction error covariance matrix of the background field is calculated as:

[0023]

[0024] In the formula, E represents mathematical expectation, and the superscript T represents transposition operation;

[0025] The Kalman gain matrix is ​​calculated based on the prediction error covariance matrix. The calculation expression is:

[0026]

[0027] In the formula, H i,j represents the observation operator, R i,j (t) represents the observation error covariance matrix, and the superscript -1 represents the matrix inverse operation;

[0028] The background field is corrected using the Kalman gain matrix to obtain the analysis field, that is, the corrected simulation state:

[0029]

[0030] The error covariance matrix of the updated analysis field is:

[0031]

[0032] Where I represents the unit matrix;

[0033] Continuously calculate and update the error covariance matrix of the ensemble Kalman filter at each time step t, and store the Kalman gain K at each time step i,j (t), constructed into the Kalman gain sequence K i,j (t1),K i,j (t2),…,K i,j (t n ).

[0034] Further, step 3 includes:

[0035] The first climate historical simulation data S i,j (t) lagged by one time period and two time periods respectively, compared with the original data sequence S i,j (t) is aligned with the Kalman gain sequence, and the data dimension is vertically expanded to form a feature matrix, which is expressed as:

[0036]

[0037] Where n is the length of the data sequence;

[0038] The Kalman gain sequence corresponding to the time is recorded as y i,j =[K i,j (3),K i,j (4),…,K i,j (n)];

[0039] Flatten all spatial grid points (i, j) and transform the characteristic matrix x of each spatial grid point i,j and the corresponding Kalman gain sequence y i,j Perform one-dimensional concatenation respectively to obtain the overall feature matrix X and the corresponding target matrix Y, which are expressed as:

[0040] X=[x 1,1 ,x 1,2 ,…,x i,j ,…]

[0041] Y=[y 1,1 ,y 1,2 ,…,y i,j ,…]

[0042] Among them, X is the first feature engineering after the historical simulation data sequence of all spatial grid points is one-dimensionalized, and Y is the target Kalman gain sequence feature engineering of all spatial grid points.

[0043] Further, step 4 includes:

[0044] The first feature engineering X and the target Kalman gain sequence feature engineering Y are input into the ResLSTM network for training. The prediction error of the Kalman gain is minimized by the back propagation algorithm. The loss function is the mean square error MSE. The training goal is to minimize the predicted Kalman gain. and the true Kalman gain sequence K i,j The root mean square error between (t) enables the ResLSTM network to accurately predict the Kalman gain of the future period; the calculation expression of the mean square error MSE is:

[0045]

[0046] Where N is the total number of spatial grid points.

[0047] Furthermore, in step 5, feature engineering is performed on the future climate prediction data, and the second feature engineering includes:

[0048] The future climate projection data P i,j (t) lagged by one time period and two time periods respectively, and compared with the original data sequence P i,j (t) Align and expand the data dimension vertically to form the feature matrix xp i,j :

[0049]

[0050] In the formula, τ is the length of the data sequence, P i,j (t) represents the future climate projection data at the (i, j)th spatial grid point at time step t;

[0051] The feature matrix of each spatial grid point is concatenated in one dimension to obtain the second feature engineering XP = [xp 1,1 ,xp 1,2 ,…,xp i,j ,…].

[0052] Further, step 6 comprises:

[0053] Calculate the average deviation of historical climate data at the spatial grid point (i, j), the expression is:

[0054]

[0055] Using mean deviation and prediction Kalman gain Correct the future climate projection data, and the calculation expression is:

[0056]

[0057] In the formula, Represents the bias-corrected future climate projection data at the spatial grid point (i, j).

[0058] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0059] 1. The present invention combines the robustness and physical mechanism of the ensemble Kalman filter bias correction method with the advantages of deep learning technology in time series prediction, and proposes a climate model bias correction method based on ensemble Kalman filtering and deep learning, which can more objectively correct the bias of historical data simulated by the global climate model and the estimated future data;

[0060] 2. The present invention enhances the physical interpretation of the bias correction process of global climate model data, effectively supporting the global future hydrological and climate scenario prediction and environmental impact assessment of climate change;

[0061] 3. The present invention can be extended to the deviation correction work of similar types of data, thus promoting the development of the deviation correction technology field. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 A flowchart of a climate model bias correction method based on ensemble Kalman filtering and deep learning;

[0063] Figure 2 A flowchart of a climate model bias correction method based on ensemble Kalman filtering and deep learning;

[0064] Figure 3 This is a schematic diagram of the structure of the ResLSTM network;

[0065] Figure 4 This is the change process of the loss function value when training the ResLSTM network;

[0066] Figure 5 It is a curve chart of the observed value, simulated value and corrected simulated value of precipitation in the training set and test set data;

[0067] Figure 6 This is the spatial distribution diagram of data in the training set and test set at different stages. DETAILED DESCRIPTION

[0068] In order to make the objectives, technical solutions and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments.

[0069] The climate model bias correction method based on ensemble Kalman filtering and deep learning described in this embodiment includes at least the following steps 1 to 6, and the flow chart is as follows: Figure 1 to Figure 2 As shown, the specific implementation process is:

[0070] Step 1: Use the global climate model to obtain the first climate historical simulation data and future climate projection data in the study area, as well as the historical observation data of the corresponding climate variables.

[0071] The Global Climate Model (GCM) is a numerical model used to simulate the Earth's climate system. GCM can be used to obtain historical simulation data and future forecast data of climate variables at spatial locations (i, j) in the study area. i and j represent the longitude and latitude of the spatial grid, respectively. Climate variables include temperature, precipitation, etc. Historical observation data of climate variables can then be obtained through data sets.

[0072] Step 2: Based on the first climate historical simulation data and historical observation data, the ensemble Mann filtering method is used to perform bias correction on the first climate historical simulation data point by point to obtain the corrected climate data sequence and Kalman gain sequence for each spatial grid point.

[0073] Furthermore, in step 2, the first climate historical simulation data is bias-corrected point by point using the ensemble Mann filtering method, and the corrected meteorological data sequence and Kalman gain sequence of each spatial grid point are obtained, including:

[0074] For each spatial grid point (i, j), the simulated state of the global climate model is represented by the state vector S i,j (t) indicates that the vector S i,j (t) represents the simulated value of the climate variable at time t, that is, the first climate historical simulation data, and the historical observation data of the corresponding climate variable is recorded as vector O i,j (t), calculate the deviation D between the historical observation data and the first climate historical simulation data at time t i,j (t) is:

[0075] D i,j (t) = O i,j (t)-S i,j (t)

[0076] In the formula, i and j represent the longitude and latitude of the spatial grid point respectively;

[0077] The ensemble Kalman filter algorithm advances the simulation state in time based on the physical equations of the climate model, initializes the ensemble Kalman filter parameters of each spatial grid point, constructs the state transfer equation and observation set with error disturbance, and obtains the predicted background field as follows:

[0078]

[0079] In the formula, ε i,j represents the simulation error, which follows a normal distribution with a mean of 0 and a variance of G, i.e., ε~N(0,G);

[0080] This background field contains the predicted state, and due to the existence of errors, the predicted state vector is uncertain. The mean of the background field is calculated, and the calculation expression is:

[0081]

[0082] In the formula, E n represents the total number of sets m in the background field;

[0083] According to the background field and its mean, the prediction error covariance matrix of the background field is calculated as:

[0084]

[0085] In the formula, E represents mathematical expectation, and the superscript T represents transposition operation;

[0086] The Kalman gain matrix is ​​calculated based on the prediction error covariance matrix. The calculation expression is:

[0087]

[0088] In the formula, H i,j represents the observation operator, which maps the state in the simulation space to the observation space; R i,j (t) represents the observation error covariance matrix, which describes the statistical characteristics of the observation error. The superscript -1 represents the matrix inverse operation.

[0089] The background field is corrected using the Kalman gain matrix to obtain the analysis field, that is, the corrected simulation state:

[0090]

[0091] The error covariance matrix of the updated analysis field is:

[0092]

[0093] Where I represents the unit matrix;

[0094] Continuously calculate and update the error covariance matrix of the ensemble Kalman filter at each time step t, and store the Kalman gain K at each time step i,j (t), constructed into the Kalman gain sequence K i,j (t1),K i,j (t2),…,K i,j (t n ). And use the analytical field at each time step Construct the corrected climate data series for each spatial grid point.

[0095] Step 3, performing hysteresis processing on the first climate historical simulation data to obtain second climate historical simulation data, performing feature engineering processing on the second climate historical simulation data and the Kalman gain sequence to obtain first feature engineering and target Kalman gain sequence feature engineering respectively;

[0096] Further, step 3 includes:

[0097] The first climate historical simulation data S i,j (t) lagged by one time period and two time periods respectively, compared with the original data sequence S i,j (t) is aligned with the Kalman gain sequence, and the data dimension is vertically expanded to form a feature matrix, which is expressed as:

[0098]

[0099] Where n is the length of the data sequence;

[0100] The Kalman gain sequence corresponding to the time is recorded as y i,j =[K i,j (3),K i,j (4),…,K i,j (n)]; feature matrix x i,j It contains simulation series of three periods (original period, lagged by one period, lagged by two periods).

[0101] Flatten all spatial grid points (i, j) and transform the characteristic matrix x of each spatial grid point i,j and the corresponding Kalman gain sequence y i,j Perform one-dimensional concatenation respectively to obtain the overall feature matrix X and the corresponding target matrix Y, which are expressed as:

[0102] X=[x 1,1 ,x 1,2 ,…,x i,j ,…]

[0103] Y=[y 1,1 ,y 1,2 ,…,y i,j ,…]

[0104] Among them, X is the first feature engineering after the historical simulation data sequence of all spatial grid points is one-dimensionalized, and Y is the target Kalman gain sequence feature engineering of all spatial grid points.

[0105] Step 4, constructing a long short-term memory network model, and training the long short-term memory network model by taking the first feature engineering as an input item and the target Kalman gain sequence feature engineering as an output item;

[0106] Further, step 4 includes:

[0107] The first feature engineering X and the target Kalman gain sequence feature engineering Y are input into the ResLSTM network for training. The prediction error of the Kalman gain is minimized by the back propagation algorithm. The loss function is the mean square error MSE. The training goal is to minimize the predicted Kalman gain. and the true Kalman gain sequence K i,j The root mean square error between (t) enables the ResLSTM network to accurately predict the Kalman gain of the future period; the calculation expression of the mean square error MSE is:

[0108]

[0109] Where N is the total number of spatial grid points.

[0110] The ResLSTM network structure used in this example is as follows Figure 3 As shown in the figure, the flow of information is controlled by a series of gating units. Each LSTM unit includes three main gate structures: input gate, forget gate and output gate. These gates determine which information should be retained and which should be discarded. The LSTM network processes the input data x at each time step t. t and the hidden state h at the previous time step t and the cell state c t .

[0111] The forget gate decides which information of the previous moment is discarded and outputs a vector with a value between 0 and 1, where 1 means it is completely retained and 0 means it is completely discarded. The mathematical model of the forget gate is:

[0112] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0113] Among them, f t represents the output data, σ represents the sigmoid activation function, W f and b f are the weight and bias of the forget gate.

[0114] The input gate determines which new information needs to be added to the memory cell. In the input gate, an activation function sigmoid layer is first used to determine which values ​​to update, and then a tanh layer is used to generate new candidate information. The mathematical model is:

[0115] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0116] Among them, i t Represents output data, W i and b i are the weights and biases of the input gate. The input gate generates new candidate memory information through the tanh layer:

[0117] c t =tanh(W c ·[h t-1 ,x t ]+b c )

[0118] Among them, Wc and b c are the weights and biases of the candidate memory cells.

[0119] Unit state update: According to the forget gate f t Determines how much memory information c of the previous time step to retain t-1 , according to the input gate i t Determine how many new candidate memory information c to add t :

[0120] c t =f t ⊙c t-1 +i t ⊙c t

[0121] The output gate determines the output state h of the current time step t First, the sigmoid layer is used to determine which parts of the memory unit information can be used as output, and then the memory unit information is modulated by the tanh function, and finally the current hidden state h is obtained. t , the mathematical model is:

[0122] o t =σ(W o ·[h t-1 ,x t ]+b o )

[0123] h t =o t ⊙tanh (c t )

[0124] Among them, W o and b o are the weight and bias terms of the output gate.

[0125] The goal of residual connection is to skip some layers and add the input directly to the output, thereby alleviating the gradient vanishing or gradient exploding problem and making the training of deep networks more stable. Assume that the original output of LSTM is h t , the input is x t , the basic form of residual connection is:

[0126]

[0127] where h res is the output after residual connection.

[0128] Step 5, feature engineering is performed on the future climate prediction data to obtain a second feature engineering, and the second feature engineering is input into the trained long short-term memory network model to obtain the Kalman gain sequence of each future period;

[0129] Furthermore, in step 5, feature engineering is performed on the future climate prediction data, and the second feature engineering includes:

[0130] Assuming that the statistical characteristics of the deviations of future projection data are consistent with the historical simulation series, the future climate projection data P i,j (t) lagged by one time period and two time periods respectively, and compared with the original data sequence P i,j (t) Align and expand the data dimension vertically to form the feature matrix xp i,j :

[0131]

[0132] In the formula, τ is the length of the data sequence, P i,j (t) represents the future climate projection data at the (i, j)th spatial grid point at time step t;

[0133] The feature matrix of each spatial grid point is concatenated in one dimension to obtain the second feature engineering XP = [xp 1,1 ,xp 1,2 ,…,xp i,j ,…].

[0134] Step 6: Based on the ensemble Kalman filter method, the future climate prediction data is bias-corrected using the average deviation of the Kalman gain sequence of each future period and the historical sequence to generate corrected future climate prediction data.

[0135] Since the future projection data and historical simulation data are output by the same GCM, it is assumed that the statistical characteristics of their deviations from actual observations are consistent, that is, the deviation D i,j (t) can be regarded as having some predictable structure or statistical properties. This deviation pattern can be used to estimate the future sequence P. i,j (t) up.

[0136] Further, step 6 comprises:

[0137] Calculate the average deviation of historical climate data at the spatial grid point (i, j), the expression is:

[0138]

[0139] Using mean deviation and prediction Kalman gain Correct the future climate projection data, and the calculation expression is:

[0140]

[0141] In the formula, Represents the bias-corrected future climate projection data at the spatial grid point (i, j).

[0142] To further verify the effectiveness and excellence of the climate model bias correction method based on ensemble Kalman filtering and deep learning described in this embodiment, the following examples are used for illustration. Taking Africa as an example, the monthly precipitation observation data from 1901 to 2015 are obtained from the CRU data set with a spatial resolution of 0.5°. The historical precipitation data simulated by GCM selects the historical simulated precipitation data output by the ACCESS_CM2 model under the r1i1p1f1 boundary condition, and the GCM output results are downscaled to 0.5° resolution using the Climatic DataOperator software using the bilinear interpolation method to obtain precipitation simulation data. After the precipitation simulation data is corrected by the correction method described in this example, precipitation correction data is obtained. The above data is divided into a training set and a test set according to the ratio, that is, the training set includes precipitation observation data, precipitation simulation data and precipitation correction data, and the test set includes precipitation observation data, precipitation simulation data and precipitation correction data.

[0143] The ResLSTM used in this example uses a 6-hidden layer network with residual connection blocks, and the structure is as follows Figure 3 As shown, the number of neurons n is 128, 64, 32, 16, 8, and 4 respectively. It is built based on TensorFlow 2.11.0 and Keras framework, the Python version is 3.10, and it runs on a server equipped with an AMD 7413 processor, Nvidia RTX A40 graphics processor, and 256GB of memory.

[0144] The constructed feature engineering X and Y are input into the ResLSTM neural network for training. The training loss function is selected as MSE, the optimizer is selected as Adam (gradient descent optimizer), the training iteration number epochs = 50, and the learning rate is initially set to 1e-3, set to 5e-4 when the iteration number reaches 20, and set to 1e-4 when the iteration number reaches 35. The loss function value during the training process is as follows Figure 4 As shown in the figure, it can be seen that the loss function value decreases steadily, the loss values ​​of the training set and the test set are consistent, and there is no obvious overfitting phenomenon.

[0145] Figure 5The observed values, simulated values, and corrected simulated values ​​of precipitation in the training set and test set data are shown. Figure (a) and Figure (b) are the time series diagrams of the training set and the test set, respectively. It can be seen from the figure that the corrected simulated data sequence is closer to the observed sequence. For the training set, the average value of the observed data is 57.5 mm, the average value of the simulated data is 60.5 mm, the Nash efficiency coefficient (NSE) of the sequence before correction is 0.50, and the relative error is 5.2%; the average value of the corrected sequence is 57.5 mm, the relative error is 0.07%, and the NSE is improved to 0.76. For the test set, the average value of the observed data is 58.2 mm, the average value of the simulated data is 62.1 mm, the relative error is 6.7%, and the NSE is 0.36; the average value of the corrected sequence is 57.7, the relative error is -0.9%, and the NSE is improved to 0.78. It can be seen that the corrected data sequence is more consistent with the actual situation obtained by observation.

[0146] Figure 6 The spatial distribution of data in the training set and test set at different stages is shown. Figure (a) and Figure (b) are the spatial distribution of precipitation observation data in the training set and test set, respectively. Figure (c) and Figure (d) are the spatial distribution of precipitation simulation data in the training set and test set, respectively. Figure (e) and Figure (f) are the spatial distribution of precipitation correction data in the training set and test set, respectively. Figure 6 It can be seen that the corrected precipitation simulation data are closer to the observed data in terms of spatial mode.

Claims

1. A climate model bias correction method based on ensemble Kalman filtering and deep learning, characterized in that: The steps include: Step 1: Use the global climate model to obtain the first climate historical simulation data and future climate projection data in the study area, and obtain the historical observation data of the corresponding climate variables; Step 2, based on the first climate historical simulation data and the historical observation data, the first climate historical simulation data is bias-corrected point by point using the ensemble Mann filtering method to obtain a corrected climate data sequence and Kalman gain sequence for each spatial grid point; Step 3, performing hysteresis processing on the first climate historical simulation data to obtain second climate historical simulation data, performing feature engineering processing on the second climate historical simulation data and the Kalman gain sequence to obtain first feature engineering and target Kalman gain sequence feature engineering respectively; Step 4, constructing a long short-term memory network model, and training the long short-term memory network model by taking the first feature engineering as an input item and the target Kalman gain sequence feature engineering as an output item; Step 5, feature engineering is performed on the future climate prediction data to obtain a second feature engineering, and the second feature engineering is input into the trained long short-term memory network model to obtain the Kalman gain sequence of each future period; Step 6: Based on the ensemble Kalman filter method, the future climate prediction data is bias-corrected using the average deviation of the Kalman gain sequence of each future period and the historical sequence to generate corrected future climate prediction data.

2. The method for correcting climate model bias based on ensemble Kalman filtering and deep learning according to claim 1, characterized in that: In step 2, the ensemble Mann filtering method is used to perform bias correction on the first climate historical simulation data point by point, and the corrected meteorological data sequence and Kalman gain sequence of each spatial grid point are obtained, including: For each spatial grid point (i, j), the simulated state of the global climate model is represented by the state vector S i,j (t) indicates that the vector S i,j (t) represents the simulated value of the climate variable at time t, that is, the first climate historical simulation data, and the historical observation data of the corresponding climate variable is recorded as vector O i,j (t), calculate the deviation D between the historical observation data and the first climate historical simulation data at time t i,j (t) is: D i,j (t)=O i,j (t)-S i,j (t) In the formula, i and j represent the longitude and latitude of the spatial grid point respectively; The ensemble Kalman filter algorithm advances the simulation state in time based on the physical equations of the climate model, constructs the state transfer equation and observation set with error disturbance, and obtains the predicted background field as follows: In the formula, ε i,j represents the simulation error, which follows a normal distribution with a mean of 0 and a variance of G, i.e., ε~N(0,G); Calculate the mean of the background field, the calculation expression is: In the formula, E n represents the total number of sets m in the background field; According to the background field and its mean, the prediction error covariance matrix of the background field is calculated as: In the formula, E represents mathematical expectation, and the superscript T represents transposition operation; The Kalman gain matrix is ​​calculated based on the forecast error covariance matrix. The calculation expression is: In the formula, H i,j represents the observation operator, R i,j (t) represents the observation error covariance matrix, and the superscript -1 represents the matrix inverse operation; The background field is corrected using the Kalman gain matrix to obtain the analysis field, that is, the corrected simulation state: The error covariance matrix of the updated analysis field is: Where I represents the unit matrix; Continuously calculate and update the error covariance matrix of the ensemble Kalman filter at each time step t, and store the Kalman gain K at each time step i,j (t), constructed into the Kalman gain sequence K i,j (t1),K i,j (t2),…,K i,j (t n ).

3. The method for correcting climate model bias based on ensemble Kalman filtering and deep learning according to claim 2, characterized in that: Step 3 includes: The first climate historical simulation data S i,j (t) lagged by one time period and two time periods respectively, compared with the original data sequence S i,j (t) is aligned with the Kalman gain sequence, and the data dimension is vertically expanded to form a feature matrix, which is expressed as: Where n is the length of the data sequence; The Kalman gain sequence corresponding to the time is recorded as y i,j =[K i,j (3),K i,j (4),…,K i,j (n)]; Flatten all spatial grid points (i, j) and transform the characteristic matrix x of each spatial grid point i,j and the corresponding Kalman gain sequence y i,j Perform one-dimensional concatenation respectively to obtain the overall feature matrix X and the corresponding target matrix Y, which are expressed as: X=[x 1,1 ,x 1,2 ,…,x i,j ,…] And=[and 1,1 ,and 1,2 ,…,and i,j ,…] Among them, X is the first feature engineering after the historical simulation data sequence of all spatial grid points is one-dimensionalized, and Y is the target Kalman gain sequence feature engineering of all spatial grid points.

4. The method for correcting climate model bias based on ensemble Kalman filtering and deep learning according to claim 3, characterized in that: Step 4 includes: The first feature engineering X and the target Kalman gain sequence feature engineering Y are input into the ResLSTM network for training. The prediction error of the Kalman gain is minimized by the back propagation algorithm. The loss function is the mean square error MSE. The training goal is to minimize the predicted Kalman gain. and the true Kalman gain sequence K i,j The root mean square error between (t) enables the ResLSTM network to accurately predict the Kalman gain of the future period; the calculation expression of the mean square error MSE is: Where N is the total number of spatial grid points.

5. The method for correcting climate model bias based on ensemble Kalman filtering and deep learning according to claim 4, characterized in that: In step 5, feature engineering is performed on the future climate projection data, and the second feature engineering includes: The future climate projection data P i,j (t) lagged by one time period and two time periods respectively, and compared with the original data sequence P i,j (t) Align and expand the data dimension vertically to form the feature matrix xp i,j : In the formula, τ is the length of the data sequence, P i,j (t) represents the future climate projection data at the (i, j)th spatial grid point at time step t; The feature matrix of each spatial grid point is concatenated in one dimension to obtain the second feature engineering XP = [xp 1,1 ,xp 1,2 ,…,xp i,j ,…].

6. The method for correcting climate model bias based on ensemble Kalman filtering and deep learning according to claim 5, characterized in that: Step 6 includes: Calculate the average deviation of historical climate data at the spatial grid point (i, j), the expression is: Using mean deviation and prediction Kalman gain Correct the future climate projection data, and the calculation expression is: In the formula, Represents the bias-corrected future climate projection data at the spatial grid point (i, j).

Images